GLM-5

Training compute

6.8×10²⁴ FLOP

Parameters

744B

Published

Feb 17, 2026

GLM-5 is an AI model developed by Z.ai (Zhipu AI) (China), first published in February 2026. It works in the language domain, on tasks such as language modeling/generation.

Training it took an estimated 6.8×10²⁴ FLOP of compute. The model has 744,000,000,000 parameters. Training ran on Huawei Ascend 910C.

Access: Open weights (unrestricted). Its weights are openly available. Epoch AI rates the confidence of this record as likely.

Full record

Organization: Z.ai (Zhipu AI)
Country of organization: China
Domain: Language
Task: Language modeling/generation
Training compute: 6.8×10²⁴ FLOP
Parameters: 744,000,000,000
Training hardware: Huawei Ascend 910C
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Epoch confidence: Likely

Benchmark results

01	Vending-Bench 2	Score	4432.1
02	WebDev Arena	Arena Score	1447.3
03	ALE Bench	Performance	765.6
04	Epoch Capabilities Index	ECI Score	146.4
05	ForecastBench	Overall score	61
06	GPQA Diamond	mean_score	0.88
07	OTIS Mock AIME 2024-2025	mean_score	0.8
08	SWE-bench Verified	mean_score	0.72
09	SimpleBench	Score (AVG@5)	0.53
10	Terminal-Bench	Accuracy mean	0.52
11	WeirdML	Accuracy	0.48
12	ARC-AGI	Score	0.45
13	CL Bench	Overall	0.19
14	Apex Agents	Pass@1 score	0.17
15	FrontierMath	mean_score	0.16
16	Posttrainbench	Average (%)	0.14
17	Chess Puzzles	mean_score	0.1
18	ARC-AGI-2	Score	0.05
19	FrontierMath Tier 4	mean_score	0.02